Click here to


Are you sure ?

Yes, do it No, cancel

Medical Image Annotation with a New Low-Rank Modeling-Based Multi-Label Active Learning Method

J Wu1, S Ruan2 , C Lian2 , S Mutic1 , M Anastasio1 , H Li1* , (1) Washington University in St. Louis, Saint Louis, MO, (2) University of Rouen, Rouen, Normandy


(Tuesday, 7/31/2018) 4:30 PM - 6:00 PM

Room: Room 207

Purpose: Most informative training examples should be annotated to train a classifier to achieve optimal performance on unseen examples. Active learning is an effective solution for example selection. However, medical image examples are always contaminated by noise with varying types and levels, which induce additional selection challenges. Most existing methods do not consider image noise independently during the selection process, which can lead to sub-optimal results. A novel low-rank modeling-based multi-label active learning (LRMMAL) method is proposed for address these issues.

Methods: The LRMMAL method is an iterative process including example informativeness calculation, example sampling, and pre-defined classifier optimization. A new example sampling strategy is developed by considering measures of example noise with example label uncertainty and label correlation for example selection. A new example noise measure strategy is designed by combining the low-rank modeling method with Chebyshev distance measure to eliminate the effect of varying noise types and patterns. A new scheme is developed to automatically determine the parameters in the sampling strategy by computing the expected cross entropy over all unlabeled examples.

Results: The LRMMAL method was compared with other four reported methods on two datasets including a total of 5930 CT images by using the k-Nearest Neighbor method as a classifier. Three measure metrics, Accuracy, Macro F1 and Micro F1, were used to evaluate the classification performance. Our method achieved superior performance on both datasets. The training examples selected by our method yield less noise level compared to those selected with other methods, which indicates the benefit by considering noise as a separate measure into the informative example selection.

Conclusion: The example noise level is effectively quantified in our method regardless of noise types and patterns. The proposed method has the potential to assist in training a high-performance classifier for medical applications.


CAD, Image Analysis


IM/TH- Image Analysis (Single modality or Multi-modality): Machine learning

Contact Email